/stable-fast

Best inference performance optimization framework for HuggingFace Diffusers on NVIDIA GPUs.

Primary LanguagePythonMIT LicenseMIT

🚀Stable Fast

wheels Upload Python Package Open In Colab

NOTE

Active development on stable-fast has been paused. I am currently working on a new torch._dynamo based project targeting new models such as stable-cascade, SD3 and Sora like mmodels. It would be faster and more flexible, as well as supporting more hardware backends rather than CUDA.

Contact is welcomed.

Discord Channel

stable-fast achieves SOTA inference performance on ALL kinds of diffuser models, even with the latest StableVideoDiffusionPipeline. And unlike TensorRT or AITemplate, which takes dozens of minutes to compile a model, stable-fast only takes a few seconds to compile a model. stable-fast also supports dynamic shape, LoRA and ControlNet out of the box.

Model torch torch.compile AIT oneflow TensorRT stable-fast
SD 1.5 (ms) 1897 1510 1158 1003 991 995
SVD-XT (s) 83 70 47

NOTE: During benchmarking, TensorRT is tested with static batch size and CUDA Graph enabled while stable-fast is running with dynamic shape.

Introduction

What is this?

stable-fast is an ultra lightweight inference optimization framework for HuggingFace Diffusers on NVIDIA GPUs. stable-fast provides super fast inference optimization by utilizing some key techniques and features:

  • CUDNN Convolution Fusion: stable-fast implements a series of fully-functional and fully-compatible CUDNN convolution fusion operators for all kinds of combinations of Conv + Bias + Add + Act computation patterns.
  • Low Precision & Fused GEMM: stable-fast implements a series of fused GEMM operators that compute with fp16 precision, which is fast than PyTorch's defaults (read & write with fp16 while compute with fp32).
  • Fused Linear GEGLU: stable-fast is able to fuse GEGLU(x, W, V, b, c) = GELU(xW + b) ⊗ (xV + c) into one CUDA kernel.
  • NHWC & Fused GroupNorm: stable-fast implements a highly optimized fused NHWC GroupNorm + Silu operator with OpenAI's Triton, which eliminates the need of memory format permutation operators.
  • Fully Traced Model: stable-fast improves the torch.jit.trace interface to make it more proper for tracing complex models. Nearly every part of StableDiffusionPipeline/StableVideoDiffusionPipeline can be traced and converted to TorchScript. It is more stable than torch.compile and has a significantly lower CPU overhead than torch.compile and supports ControlNet and LoRA.
  • CUDA Graph: stable-fast can capture the UNet, VAE and TextEncoder into CUDA Graph format, which can reduce the CPU overhead when the batch size is small. This implemention also supports dynamic shape.
  • Fused Multihead Attention: stable-fast just uses xformers and makes it compatible with TorchScript.

My next goal is to keep stable-fast as one of the fastest inference optimization frameworks for diffusers and also provide both speedup and VRAM reduction for transformers. In fact, I already use stable-fast to optimize LLMs and achieve a significant speedup. But I still need to do some work to make it more stable and easy to use and provide a stable user interface.

Differences With Other Acceleration Libraries

  • Fast: stable-fast is specialy optimized for HuggingFace Diffusers. It achieves a high performance across many libraries. And it provides a very fast compilation speed within only a few seconds. It is significantly faster than torch.compile, TensorRT and AITemplate in compilation time.
  • Minimal: stable-fast works as a plugin framework for PyTorch. It utilizes existing PyTorch functionality and infrastructures and is compatible with other acceleration techniques, as well as popular fine-tuning techniques and deployment solutions.
  • Maximum Compatibility: stable-fast is compatible with all kinds of HuggingFace Diffusers and PyTorch versions. It is also compatible with ControlNet and LoRA. And it even supports the latest StableVideoDiffusionPipeline out of the box!

Installation

NOTE: stable-fast is currently only tested on Linux and WSL2 in Windows. You need to install PyTorch with CUDA support at first (versions from 1.12 to 2.1 are suggested).

I only test stable-fast with torch>=2.1.0, xformers>=0.0.22 and triton>=2.1.0 on CUDA 12.1 and Python 3.10. Other versions might build and run successfully but that's not guaranteed.

Install Prebuilt Wheels

Download the wheel corresponding to your system from the Releases Page and install it with pip3 install <wheel file>.

Currently both Linux and Windows wheels are available.

# Change cu121 to your CUDA version and <wheel file> to the path of the wheel file.
# And make sure the wheel file is compatible with your PyTorch version.
pip3 install --index-url https://download.pytorch.org/whl/cu121 \
    'torch>=2.1.0' 'xformers>=0.0.22' 'triton>=2.1.0' 'diffusers>=0.19.3' \
    '<wheel file>'

Install From Source

# Make sure you have CUDNN/CUBLAS installed.
# https://developer.nvidia.com/cudnn
# https://developer.nvidia.com/cublas

# Install PyTorch with CUDA and other packages at first.
# Windows user: Triton might be not available, you could skip it.
# NOTE: 'wheel' is required or you will meet `No module named 'torch'` error when building.
pip3 install wheel 'torch>=2.1.0' 'xformers>=0.0.22' 'triton>=2.1.0' 'diffusers>=0.19.3'

# (Optional) Makes the build much faster.
pip3 install ninja

# Set TORCH_CUDA_ARCH_LIST if running and building on different GPU types.
# You can also install the latest stable release from PyPI.
# pip3 install -v -U stable-fast
pip3 install -v -U git+https://github.com/chengzeyi/stable-fast.git@main#egg=stable-fast
# (this can take dozens of minutes)

NOTE: Any usage outside sfast.compilers is not guaranteed to be backward compatible.

NOTE: To get the best performance, xformers and OpenAI's triton>=2.1.0 need to be installed and enabled. You might need to build xformers from source to make it compatible with your PyTorch.

Usage

Optimize StableDiffusionPipeline

stable-fast is able to optimize StableDiffusionPipeline and StableDiffusionPipelineXL directly.

import time
import torch
from diffusers import (StableDiffusionPipeline,
                       EulerAncestralDiscreteScheduler)
from sfast.compilers.diffusion_pipeline_compiler import (compile,
                                                         CompilationConfig)

def load_model():
    model = StableDiffusionPipeline.from_pretrained(
        'runwayml/stable-diffusion-v1-5',
        torch_dtype=torch.float16)

    model.scheduler = EulerAncestralDiscreteScheduler.from_config(
        model.scheduler.config)
    model.safety_checker = None
    model.to(torch.device('cuda'))
    return model

model = load_model()

config = CompilationConfig.Default()
# xformers and Triton are suggested for achieving best performance.
try:
    import xformers
    config.enable_xformers = True
except ImportError:
    print('xformers not installed, skip')
try:
    import triton
    config.enable_triton = True
except ImportError:
    print('Triton not installed, skip')
# CUDA Graph is suggested for small batch sizes and small resolutions to reduce CPU overhead.
# But it can increase the amount of GPU memory used.
# For StableVideoDiffusionPipeline it is not needed.
config.enable_cuda_graph = True

model = compile(model, config)

kwarg_inputs = dict(
    prompt=
    '(masterpiece:1,2), best quality, masterpiece, best detailed face, a beautiful girl',
    height=512,
    width=512,
    num_inference_steps=30,
    num_images_per_prompt=1,
)

# NOTE: Warm it up.
# The initial calls will trigger compilation and might be very slow.
# After that, it should be very fast.
for _ in range(3):
    output_image = model(**kwarg_inputs).images[0]

# Let's see it!
# Note: Progress bar might work incorrectly due to the async nature of CUDA.
begin = time.time()
output_image = model(**kwarg_inputs).images[0]
print(f'Inference time: {time.time() - begin:.3f}s')

# Let's view it in terminal!
from sfast.utils.term_image import print_image

print_image(output_image, max_width=80)

Refer to examples/optimize_stable_diffusion_pipeline.py for more details.

You can check this Colab to see how it works on T4 GPU: Open In Colab

Optimize LCM Pipeline

stable-fast is able to optimize the newest latent consistency model pipeline and achieve a significant speedup.

Refer to examples/optimize_lcm_pipeline.py for more details about how to optimize normal SD model with LCM LoRA. Refer to examples/optimize_lcm_pipeline.py for more details about how to optimize the standalone LCM model.

Optimize StableVideoDiffusionPipeline

stable-fast is able to optimize the newest StableVideoDiffusionPipeline and achieve a 2x speedup

Refer to examples/optimize_stable_video_diffusion_pipeline.py for more details

Dynamically Switch LoRA

Switching LoRA dynamically is supported but you need to do some extra work. It is possible because the compiled graph and CUDA Graph share the same underlaying data (pointers) with the original UNet model. So all you need to do is to update the original UNet model's parameters inplace.

The following code assumes you have already load a LoRA and compiled the model, and you want to switch to another LoRA.

If you don't enable CUDA graph and keep preserve_parameters = True, things could be much easier. The following code might not even be needed.

# load_state_dict with assign=True requires torch >= 2.1.0

def update_state_dict(dst, src):
    for key, value in src.items():
        # Do inplace copy.
        # As the traced forward function shares the same underlaying data (pointers),
        # this modification will be reflected in the traced forward function.
        dst[key].copy_(value)

# Switch "another" LoRA into UNet
def switch_lora(unet, lora):
    # Store the original UNet parameters
    state_dict = unet.state_dict()
    # Load another LoRA into unet
    unet.load_attn_procs(lora)
    # Inplace copy current UNet parameters to the original unet parameters
    update_state_dict(state_dict, unet.state_dict())
    # Load the original UNet parameters back.
    # We use assign=True because we still want to hold the references
    # of the original UNet parameters
    unet.load_state_dict(state_dict, assign=True)

switch_lora(compiled_model.unet, lora_b_path)

Model Quantization

stable-fast extends PyTorch's quantize_dynamic functionality and provides a dynamically quantized linear operator on CUDA backend. By enabling it, you could get a slight VRAM reduction for diffusers and significant VRAM reduction for transformers, and cound get a potential speedup (not always).

For SD XL, it is expected to see VRAM reduction of 2GB with an image size of 1024x1024.

def quantize_unet(m):
    from diffusers.utils import USE_PEFT_BACKEND
    assert USE_PEFT_BACKEND
    m = torch.quantization.quantize_dynamic(m, {torch.nn.Linear},
                                            dtype=torch.qint8,
                                            inplace=True)
    return m

model.unet = quantize_unet(model.unet)
if hasattr(model, 'controlnet'):
    model.controlnet = quantize_unet(model.controlnet)

Refer to examples/optimize_stable_diffusion_pipeline.py for more details.

Some Common Methods To Speed Up PyTorch

# TCMalloc is highly suggested to reduce CPU overhead
# https://github.com/google/tcmalloc
LD_PRELOAD=/path/to/libtcmalloc.so python3 ...
import packaging.version
import torch

if packaging.version.parse(torch.__version__) >= packaging.version.parse('1.12.0'):
    torch.backends.cuda.matmul.allow_tf32 = True

Performance Comparison

Performance varies very greatly across different hardware/software/platform/driver configurations. It is very hard to benchmark accurately. And preparing the environment for benchmarking is also a hard job. I have tested on some platforms before but the results may still be inaccurate. Note that when benchmarking, the progress bar showed by tqdm may be inaccurate because of the asynchronous nature of CUDA. To solve this problem, I use CUDA Event to measure the speed of iterations per second accurately.

stable-fast is expected to work better on newer GPUs and newer CUDA versions. On older GPUs, the performance increase might be limited. During benchmarking, the progress bar might work incorrectly because of the asynchronous nature of CUDA.

RTX 4080 (512x512, batch size 1, fp16, in WSL2)

This is my personal gaming PC😄. It has a more powerful CPU than those from cloud server providers.

Framework SD 1.5 SD XL (1024x1024) SD 1.5 ControlNet
Vanilla PyTorch (2.1.0) 29.5 it/s 4.6 it/s 19.7 it/s
torch.compile (2.1.0, max-autotune) 40.0 it/s 6.1 it/s 21.8 it/s
AITemplate 44.2 it/s
OneFlow 53.6 it/s
AUTO1111 WebUI 17.2 it/s 3.6 it/s
AUTO1111 WebUI (with SDPA) 24.5 it/s 4.3 it/s
TensorRT (AUTO1111 WebUI) 40.8 it/s
TensorRT Official Demo 52.6 it/s
stable-fast (with xformers & Triton) 51.6 it/s 9.1 it/s 36.7 it/s

H100

Thanks for @Consceleratus and @harishp's help, I have tested speed on H100.

Framework SD 1.5 SD XL (1024x1024) SD 1.5 ControlNet
Vanilla PyTorch (2.1.0) 54.5 it/s 14.9 it/s 35.8 it/s
torch.compile (2.1.0, max-autotune) 66.0 it/s 18.5 it/s
stable-fast (with xformers & Triton) 104.6 it/s 21.6 it/s 72.6 it/s

A100

Thanks for @SuperSecureHuman and @jon-chuang's help, benchmarking on A100 is available now.

Framework SD 1.5 SD XL (1024x1024) SD 1.5 ControlNet
Vanilla PyTorch (2.1.0) 35.6 it/s 8.7 it/s 25.1 it/s
torch.compile (2.1.0, max-autotune) 41.9 it/s 10.0 it/s
stable-fast (with xformers & Triton) 61.8 it/s 11.9 it/s 41.1 it/s

Compatibility

Model Supported
Hugging Face Diffusers (1.5/2.1/XL) Yes
With ControlNet Yes
With LoRA Yes
Latent Consistency Model Yes
SDXL Turbo Yes
Stable Video Diffusion Yes
Functionality Supported
Dynamic Shape Yes
Text to Image Yes
Image to Image Yes
Image Inpainting Yes
UI Framework Supported Link
AUTOMATIC1111 WIP
SD Next Yes SD Next
ComfyUI Yes ComfyUI_stable_fast
Operating System Supported
Linux Yes
Windows Yes
Windows WSL Yes

Troubleshooting

Refer to doc/troubleshooting.md for more details.

And you can join the Discord Channel to ask for help.